Why big(ger) data?

Generalization

Robustness

Variability

Consistency



  • Michael Frank
  • Mika Braginsky
  • Dan Yurovsky
  • Virginia Marchman

CDI

  • Parent report checklist
  • Words and Gestures (8–18 months)
    Words and Sentences (16–30 months)
  • Relatively easy to collect
  • Reliable and valid
  • Adapted to many languages

Wordbank

Wordbank

library(tidyverse)
library(ggthemes)
library(wordbankr)
ls("package:wordbankr")
## [1] "fit_aoa"                 "get_administration_data"
## [3] "get_crossling_data"      "get_crossling_items"    
## [5] "get_instrument_data"     "get_instruments"        
## [7] "get_item_data"           "get_sources"            
## [9] "summarise_items"

Wordbank: instruments

intruments <- get_instruments()
intruments

Wordbank: sources

sources <- get_sources()
sources

Wordbank: sources

sources %>% distinct(instrument_language, instrument_form)

Wordbank: administrations

admins_eng_ws <- get_administration_data(language = "English (American)", form = "WS")
admins_eng_ws
n_distinct(admins_eng_ws$data_id)
## [1] 5520

Wordbank: administrations

admins_eng_ws %>% count(age)

Wordbank: administrations

ggplot(admins_eng_ws, aes(x = age, y = production)) +
  geom_jitter(colour = "grey", size = 0.5) +
  geom_smooth()

Wordbank: administrations

admins_russian <- get_administration_data(language = "Russian")
admins_russian

Wordbank: administrations

admins_ws <- get_administration_data(form = "WS")
admins_ws

Wordbank: administrations

admins <- get_administration_data()
admins
nrow(admins)
## [1] 81973

Wordbank: administrations

admins %>% count(language, form)

Wordbank: items

items_eng_ws <- get_item_data(language = "English (American)", form = "WS")
items_eng_ws

Wordbank: items

items_eng_ws %>% distinct(type)

Wordbank: items

items_eng_ws %>% distinct(category)

Wordbank: items

items_eng_ws %>% distinct(lexical_category)

Wordbank: items

items <- get_item_data()
items

Wordbank: items

items %>% count(language, form)

Wordbank: data

items_eng_ws %>% filter(definition == "dog")
get_instrument_data(language = "English (American)", form = "WS",
                    items = "item_26")

Wordbank: data

get_instrument_data(language = "English (American)", form = "WS",
                    items = "item_26", administrations = TRUE, iteminfo = TRUE)

Wordbank: data

twos <- admins_eng_ws %>% filter(age == 24)
dog <- items_eng_ws %>% filter(definition == "dog")
get_instrument_data(language = "English (American)", form = "WS",
                    items = "item_26", administrations = twos, iteminfo = dog)

Wordbank

Exercises

  1. Compute and plot median productive vocabulary size (as proportion of total words) over age in each language. Limit to WS data for children 16-30 months old.

  2. For English WS data, compute and plot the proportion of children that produce each word in the “toys” category at each age. Now do the same thing but separately for girls and boys.

Wordbank

Median vocabulary sizes

num_words <- items %>%
  filter(form == "WS", type == "word") %>%
  group_by(language) %>%
  summarise(words = n())
vocab_summary <- admins %>%
  filter(form == "WS", age >= 16, age <= 30) %>%
  left_join(num_words) %>%
  mutate(prop_vocab = production / words) %>%
  group_by(language, age) %>%
  summarise(median_vocab = median(prop_vocab))
ggplot(vocab_summary, aes(x = age, y = median_vocab)) +
  facet_wrap(~language) +
  geom_point() +
  ylim(0, 1) +
  labs(x = "Age (months)", y = "Productive vocabulary size")

Wordbank

Toy trajectories

toys <- items_eng_ws %>%
  filter(type == "word", category == "toys")

toys_data <- get_instrument_data(language = "English (American)", form = "WS",
                                 items = toys$item_id,
                                 administrations = admins_eng_ws,
                                 iteminfo = toys) %>%
  mutate(produces = !is.na(value) & value == "produces")
toys_summary <- toys_data  %>%
  group_by(definition, age) %>%
  summarise(total = n(),
            prop_produces = sum(produces) / total)
ggplot(toys_summary, aes(x = age, y = prop_produces)) +
  facet_wrap(~definition) +
  geom_smooth(aes(weight = total), method = "glm", se = FALSE,
              method.args = list(family = "binomial")) +
  labs(x = "Age (months)", y = "Proportion of children producing")

Wordbank

Toy trajectories by sex

toys_summary_sex <- toys_data %>%
  filter(!is.na(sex)) %>%
  group_by(definition, age, sex) %>%
  summarise(total = n(),
            prop_produces = sum(produces) / total)
ggplot(toys_summary_sex, aes(x = age, y = prop_produces, colour = sex)) +
  facet_wrap(~definition) +
  geom_smooth(aes(weight = total), method = "glm", se = FALSE,
              method.args = list(family = "binomial")) +
  scale_colour_ptol(name = "") +
  labs(x = "Age (months)", y = "Proportion of children producing") +
  theme(legend.position = "top")

Wordbank: AoA

Wordbank: AoA

Wordbank: AoA

Wordbank: AoA

fit_aoa(toys_data) %>% select(definition, aoa)
fit_aoa(toys_data, method = "glmrob", proportion = 0.8) %>% select(definition, aoa)

Wordbank: unilemmas

get_crossling_items()
get_crossling_data(uni_lemmas = "dog")

childes-db



  • Michael Frank
  • Stephan Meylan
  • Alessandro Sanchez
  • Mika Braginsky
  • Dan Yurovsky
  • Kyle MacDonald

CHILDES

Transcriptions of child and child-directed speech

@Participants:  CHI Shem Target_Child , MOT Mother , INV Cindy Investigator
@ID:    eng|Clark|CHI|2;2.16|male|typical|UC|Target_Child|||
@ID:    eng|Clark|MOT|||||Mother|||
@ID:    eng|Clark|INV|||||Investigator|||
@Date:  21-APR-1976
@Comment:   Clark (NSF SOC 75-17126) , SESSION ONE
*CHI:   yeah .
%mor:   co|yeah .
%gra:   1|0|INCROOT 2|1|PUNCT
*INV:   what was the book about ?
%mor:   pro:wh|what cop|be&PAST&13S art|the n|book adv|about ?
%gra:   1|2|LINK 2|0|ROOT 3|2|PRED 4|3|OBJ 5|4|NJCT 6|2|PUNCT
*CHI:   ow .
%mor:   co|ow .
%gra:   1|0|INCROOT 2|1|PUNCT
*INV:   ow ?
%mor:   co|ow ?
%gra:   1|0|INCROOT 2|1|PUNCT
*CHI:   yeah .
%mor:   co|yeah .
%gra:   1|0|INCROOT 2|1|PUNCT

childes-db

childes-db

library(childesr)
ls("package:childesr")
##  [1] "connect_to_childes"     "get_collections"       
##  [3] "get_contexts"           "get_corpora"           
##  [5] "get_database_version"   "get_participants"      
##  [7] "get_speaker_statistics" "get_tokens"            
##  [9] "get_transcripts"        "get_types"             
## [11] "get_utterances"
get_database_version()
## [1] "2018.1"

childes-db: collections, corpora

get_collections()
get_corpora()

childes-db: transcripts

get_transcripts(collection = "Eng-NA")
get_transcripts(corpus = c("Brown", "Clark"))

childes-db: participants

get_participants(corpus = "Clark")
get_participants(collection = "Eng-NA", age = c(24, 36))

childes-db: utterances

get_utterances(corpus = "Clark", role = "target_child")
get_utterances(corpus = "Clark", role_exclude = "target_child")

childes-db: types

get_types(corpus = "Clark", type = "dog")
get_types(collection = "Eng-NA", role = "target_child", type = "dog")

childes-db: tokens

get_tokens(corpus = "Clark", role = "target_child", token = "dog")
get_tokens(corpus = "Clark", role = "target_child", token = "dog", replace = FALSE)

childes-db: tokens

get_tokens(corpus = "Clark", role = "target_child", token = c("dog", "cat"))
get_tokens(corpus = "Clark", role = "target_child", token = "dog%")

childes-db: tokens

get_tokens(corpus = "Clark", role = "target_child", token = "*", stem = "run")
get_tokens(corpus = "Clark", role = "target_child", token = "*", part_of_speech = "v")

childes-db: speaker statistics

brown_stats <- get_speaker_statistics(corpus = "Brown", role = "target_child")
brown_stats

childes-db: speaker statistics

ggplot(brown_stats, aes(x = target_child_age, y = mlu_w, colour = target_child_name)) +
  geom_point() +
  geom_smooth(method = "lm", se = FALSE) +
  scale_colour_ptol(name = "")

childes-db

Exercises

  1. For each corpus, plot the number of transcripts in it against its mean length in number of tokens spoken by everyone other than the target child.

  2. Retrieve and plot the number of times each child in the Brown corpus said each inflection of the verb “go” over age.

childes-db

Transcript stats

transcripts <- get_transcripts()
corpus_transcripts <- transcripts %>%
  group_by(corpus_id, corpus_name, language) %>%
  summarise(num_transcripts = n())
speaker_stats <- get_speaker_statistics(role_exclude = "target_child")
corpus_tokens <- speaker_stats %>%
  group_by(corpus_id) %>%
  summarise(mean_tokens = mean(num_tokens)) %>%
  left_join(corpus_transcripts)
ggplot(corpus_tokens, aes(x = num_transcripts, y = mean_tokens)) +
  geom_point()

childes-db

Transcript stats

childes-db

Brown “go” frequencies

go_tokens <- get_tokens(corpus = "Brown", role = "target_child",
                        stem = "go", token = "*")
go_summary <- go_tokens %>%
  mutate(age = cut(target_child_age, 20)) %>%
  group_by(target_child_name, age, gloss) %>%
  summarise(num_tokens = n())
ggplot(go_summary, aes(x = age, y = num_tokens, colour = gloss)) +
  facet_wrap(~target_child_name, scales = "free") +
  geom_line(aes(group = gloss)) +
  scale_colour_ptol()

childes-db

Brown “go” frequencies

Mini-project!

Use data from Wordbank and/or childes-db to explore some question about language learning. Here’s a few ideas:

Wordbank
– Explore the relationship between vocabulary size and grammar ability (the items of type complexity).
– Look at the composition of vocabulary – what proportion of words that children know are which lexical category – and how it changes over age.

childes-db
– Characterize the developmental trajectory of children’s lexical diversity (e.g. MTLD) and how it differs by gender.
– Estimate the frequencies of color terms (or some other interesting set of words) in speech to children over age.

Both
– For some set of words, estimate their age of acquisition from Wordbank and frequency in child-directed speech from childes-db and examine the relationship between them.
– Determine which words are earliest-learned according to CDI data and according to corpus data and compare the two.

Resources

Wordbank
– wordbank.stanford.edu
– github.com/langcog/wordbankr
– langcog.github.io/wordbankr
– mb-cdi.stanford.edu
Citation: Frank, M. C., Braginsky, M., Yurovsky, D., & Marchman, V. A. (2017). Wordbank: An open repository for developmental vocabulary data. Journal of Child Language, 44(3), 677-694.

childes-db
– childes-db.stanford.edu/
– github.com/langcog/childesr
– childes.talkbank.org
Citation: Sanchez, A., Meylan, S. C., Braginsky, M., MacDonald, K. E., Yurovsky, D., & Frank, M. C. (2019). childes-db: A flexible and reproducible interface to the Child Language Data Exchange System. Behavior Research Methods, 1-14.

This presentation
– github.com/mikabr/acq-tools
– mikabr.github.io/acq-tools

mikabr@mit.edu